ResSepDTAF: improving protein-ligand binding affinity predictions with residual separable convolutions and sequence-structural integration

Prediction of protein-ligand binding affinity holds significant guidance value in fields such as drug development and protein molecular design. This study proposes a novel deep learning model, ResSepDTAF, for predicting protein-ligand binding affinity. The model integrates depthwise separable convolutions and residual modules, employing supervised learning for regression prediction. We utilize the raw feature vectors of proteins and small molecules, processed through convolutional and fully connected layers to predict affinity values. Experimental results show that ResSepDTAF achieves an RMSE of 1.209, significantly lower than Pafnucy (1.418), DeepDTAF (1.355), and DeepAtom (1.318), indicating higher prediction accuracy. Additionally, the correlation coefficient R for ResSepDTAF is improved by 17.9%, 15.8%, and 13.2% over Pafnucy, DeepDTAF, and DeepAtom, respectively. Other metrics such as MAE, SD, and CI also outperform those of other models. Overall, ResSepDTAF demonstrates excellent performance in predicting protein-ligand binding affinity and shows high potential for practical applications.